Cache Invalidation Race

[tustvold]

Currently in various places we cache connections, and invalidate them on error. This is currently racy, and may end up invalidating a connection different from the one that error-ed, if something else concurrently invalidated the cache and then fetched a new connection.

This is unlikely to be a problem in practice, but the fix is simple - pass the BrokerConnection to the invalidate call and only invalidate if Arc::ptr_eq is true